feat: Add support for DPO

[sandeepchittilla]

Closes #504

This PR adds Direct Policy Optimization as introduced in https://arxiv.org/abs/2305.18290

Loss calculation and concatenated forward pass implementations are adapted from the original TRL library

[sandeepchittilla]

The WANDB job : https://wandb.ai/sharma-sandeepch/trlx/runs/f7ym4m9y?workspace=user-sharma-sandeepch

(updated the link to point to a run with a larger batch size)

[sandeepchittilla]

@PhungVanDuy @maxreciprocate I cannot seem to request a review (probably due to permission issues / first contribution reasons). Could you please advise?

[PhungVanDuy]

@PhungVanDuy @maxreciprocate I cannot seem to request a review (probably due to permission issues / first contribution reasons). Could you please advise?

Thank you so much for the great PR. We will review this PR asap! Can you share any wandb that you ran?

[sandeepchittilla]

Thank you so much @PhungVanDuy for reviewing 🙏

Yes it's the same wandb run i shared above. Here you go : https://wandb.ai/sharma-sandeepch/trlx/runs/f7ym4m9y?workspace=user-sharma-sandeepch

[LouisCastricato]

Any update?

[sandeepchittilla]

@PhungVanDuy @maxreciprocate I cannot seem to request a review (probably due to permission issues / first contribution reasons). Could you please advise?

Thank you so much for the great PR. We will review this PR asap! Can you share any wandb that you ran?

@PhungVanDuy were you able to review this?

[PhungVanDuy]

@PhungVanDuy @maxreciprocate I cannot seem to request a review (probably due to permission issues / first contribution reasons). Could you please advise?

Thank you so much for the great PR. We will review this PR asap! Can you share any wandb that you ran?

@PhungVanDuy were you able to review this?

I saw your wandb but actually the chart quite messup, seems reward/accuracies and reward/margin not clearly increase. I guess because you used gpt2 instead of an SFT model on HH to do DPO. Can you use this SFT model and this preference dataset to train with this branch?

[sandeepchittilla]

@PhungVanDuy @maxreciprocate I cannot seem to request a review (probably due to permission issues / first contribution reasons). Could you please advise?

Thank you so much for the great PR. We will review this PR asap! Can you share any wandb that you ran?

@PhungVanDuy were you able to review this?

I saw your wandb but actually the chart quite messup, seems reward/accuracies and reward/margin not clearly increase. I guess because you used gpt2 instead of an SFT model on HH to do DPO. Can you use this SFT model and this preference dataset to train with this branch?

That's indeed what I did for a quick iteration and because I was limited on the compute i had. I will run it with the mistral-7b on the ultrafeedback dataset and get back asap.

[sandeepchittilla]

@PhungVanDuy sorry for the delay, the gpus aren't always available. Here is a dpo run (ongoing) of 1 epoch with mistral-7b-sft-beta on the ultrafeedback_binarized dataset : https://wandb.ai/sharma-sandeepch/trlx/runs/kfpmeonf?workspace=user-sharma-sandeepch

Note :

• Ultrafeedback is a challenging dataset for DPO because the rejected responses are randomly sampled
• I have not done a sft pass on the data so we see some fluctuating plots.
• I have limited memory and GPUs are not the best in class so I've chosen only a subset of test_prefs for evaluation

[PhungVanDuy]

@PhungVanDuy sorry for the delay, the gpus aren't always available. Here is a dpo run (ongoing) of 1 epoch with mistral-7b-sft-beta on the ultrafeedback_binarized dataset : https://wandb.ai/sharma-sandeepch/trlx/runs/kfpmeonf?workspace=user-sharma-sandeepch

Note :

• Ultrafeedback is a challenging dataset for DPO because the rejected responses are randomly sampled
• I have not done a sft pass on the data so we see some fluctuating plots.
• I have limited memory and GPUs are not the best in class so I've chosen only a subset of test_prefs for evaluation

Thank you for your information, I will use SFT-beta, to check this. Let me help you to run on my cluster.

[PhungVanDuy]

@PhungVanDuy sorry for the delay, the gpus aren't always available. Here is a dpo run (ongoing) of 1 epoch with mistral-7b-sft-beta on the ultrafeedback_binarized dataset : https://wandb.ai/sharma-sandeepch/trlx/runs/kfpmeonf?workspace=user-sharma-sandeepch Note :

• Ultrafeedback is a challenging dataset for DPO because the rejected responses are randomly sampled
• I have not done a sft pass on the data so we see some fluctuating plots.
• I have limited memory and GPUs are not the best in class so I've chosen only a subset of test_prefs for evaluation

Thank you for your information, I will use SFT-beta, to check this. Let me help you to run on my cluster.

@sandeepchittilla can you add my discord with the handle: duyphung.ai, it will be easier to discuss on this. Thank you so much.

[StellaAthena]

I'm excited about DPO support and I hope it'll be added soon!